See axolotl config

axolotl version: 0.9.2

base_model: Qwen/Qwen3-4B
# Automatically upload checkpoint and final model to HF
hub_model_id: Rexhaif/Qwen3-4B-MTEval-SFT
hub_private_repo: false


load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: tokenizer_default
datasets:
  - path: Rexhaif/wmt23-pairs-sft
    split: "train"
    type: chat_template
    field_messages: messages
    roles_to_train: ["assistant"]

shuffle_merged_datasets: true

skip_prepare_dataset: false
dataset_prepared_path: ./data/wmt23-pairs-sft
output_dir: /hnvme/workspace/v106be28-outputs/sft-4b

dataloader_prefetch_factor: 32
dataloader_num_workers: 2
dataloader_pin_memory: true

gc_steps: 1

sequence_len: 512
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: false

wandb_project: llm-reasoning-mt-eval
wandb_entity:
wandb_name: qwen3-4b-sft

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true
gradient_accumulation_steps: 2
micro_batch_size: 32  # should match num_generations / num_gpus

optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 5.0e-5
cosine_min_lr_ratio: 1.0e-7
max_grad_norm: 1.0
weight_decay: 0.1

bf16: true
tf32: true

flash_attention: true
flash_attn_fuse_qkv: true
flash_attn_fuse_mlp: true
auto_resume_from_checkpoints: true

n_epochs: 3
logging_steps: 10
warmup_ratio: 0.1
evals_per_epoch: 10
saves_per_epoch: 10
save_total_limit: 1
#max_steps: 5000
seed: 42
val_set_size: 0.01

gradient_checkpointing: false
gradient_checkpointing_kwargs:
  use_reentrant: false

Qwen3-4B-MTEval-SFT

This model is a fine-tuned version of Qwen/Qwen3-4B on the Rexhaif/wmt23-pairs-sft dataset. It achieves the following results on the evaluation set:

• Loss: 0.0511

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 256
• total_eval_batch_size: 128
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 101
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0010	1	19.7622
0.251	0.1003	102	0.2284
0.2008	0.2007	204	0.1928
0.1571	0.3010	306	0.1638
0.1264	0.4014	408	0.1307
0.0964	0.5017	510	0.1090
0.0933	0.6021	612	0.0939
0.0628	0.7024	714	0.0762
0.0581	0.8028	816	0.0598
0.0519	0.9031	918	0.0511

Framework versions

• Transformers 4.51.3
• Pytorch 2.6.0+cu124
• Datasets 3.5.1
• Tokenizers 0.21.1